“A major step-function improvement in large language models occurred at the beginning of last year with the introduction of a "thinking" process, where models reason internally before providing an answer.”